Approve-with-comments re-review; no blockers. All 7 actionable points (8 is a
forward-looking architecture note — recommendation A, keep as-is):
1. chat-markdown.util spec: restore parity coverage of the removed client spec —
tool error state (+ errorText), unknown-tool fallback (`Ran tool <name>` en /
`Выполнил инструмент <name>` ru), and the circular-output stringify catch.
2. findAllByChat row cap is now testable (injectable limit) + an int-spec proves
truncation on a modest volume.
3. Stability: the per-step durability updates are SERIALIZED via a promise chain
(stepUpdateChain) so they commit in step order — onlyIfStreaming already
closed the finalize race, this closes inter-step ordering.
4. findAllByChat keeps the NEWEST messages on truncation (order DESC + reverse,
like findRecent) and logs a warning with chatId, instead of silently dropping
the newest tail.
5. The LABELS parity comment already references the real path (tool-parts.tsx /
toolLabelKey) — confirmed accurate.
6. Removed the redundant 'off-by-one boundary' test (strict subset of the two
adjacent prepareAgentStep cases).
7. Extracted the terminal-finalize dispatch into a shared `applyFinalize`, used
by BOTH the service's finalizeAssistant and its test — the test now exercises
the real path, not a copy, so a production drift fails it.
Verified: server build + 325 ai-chat unit + 6 integration; prettier clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
15-point review of the persistent-history PR. Architecture decisions: crash
recovery = recency threshold; tool-label duplication = leave as-is.
Must-fix:
1. Boot-sweep bounded by recency. sweepStreaming now also requires
`updatedAt < now() - SWEEP_STREAMING_STALE_MS` (10 min), so a fresh replica's
startup sweep can't abort a turn another replica is actively streaming
(multi-instance deploy). Int-spec: a FRESH 'streaming' row is NOT swept, a
STALE one IS.
2. Restore export during the FIRST streaming turn of a new chat (#174). The
server chatId is now adopted EARLY (in-place, on the start-chunk metadata) via
a new `onServerChatId` callback wired through use-chat-session → chat-thread,
so `activeChatId` is set at turn start and the Copy button is live mid-first-
turn (canExport = !!activeChatId). Hook tests for early/in-place/no-op adopt.
3. Cover finalizeAssistant's fallback-insert branch: extracted pure
`planFinalizeAssistant(assistantId)` (update when id present, insert when the
upfront insert failed) + a dispatch harness test for both arms.
Tests: onModuleInit lifecycle spec (sweep called; throw → resolves + warns);
int-spec updatedAt assertion → toBeGreaterThan.
Cleanups: cap findAllByChat at 5000 rows; upfront-insert-failure log carries
chatId+workspaceId; removed the now-dead buildPartialAssistantRecord (only the
spec consumed it; shapes still pinned by the flushAssistant suite); controller
passes `lang: dto.lang` (normalizeLang handles undefined); dropped a no-op
`?? undefined` in errorOf; documented the content-column semantics change
(concatenated step text, UI renders from metadata.parts); CHANGELOG [Unreleased]
entry (#183, #174); reworded the stale LABELS parity comment.
Verified: server build + 323 ai-chat unit + 5 integration; client tsc + 160
ai-chat unit; prettier clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
8-point multi-aspect review of the batch PR; security/regressions were clean.
1. Lease leak: the #180 reorder moved `toolsFor` (which leases external MCP
clients, refCount+1) ahead of buildSystemPrompt + forUser, but the only
release (closeExternalClients) was bound to the streamText callbacks. A throw
in between leaked the lease (refCount stuck, undici sockets held until
restart). Define closeExternalClients right after the lease and wrap
buildSystemPrompt+forUser in try/catch that closes-then-rethrows.
2. Cover the patch_node/delete_node dup-id refusal (#159#6): extract the guard
into a pure `assertUnambiguousMatch` (node-ops) and unit-test 0/1/>1.
3. Regress the body-before-title order (#159#10): mock-HTTP test (collab fails
fast against a server with no WS upgrade) asserts /pages/update (title) is
NEVER posted when the body write fails — for updatePage AND updatePageJson.
4. CHANGELOG [Unreleased]: #180, #168 (Added); #163 (Fixed).
5. Add the missing en-US i18n keys (Back to references / {{label}}).
6. Drop the duplicate content/empty/blank cases in ai-chat.prompt.spec.ts (they
repeat the buildMcpToolingBlock unit tests); keep only sandwich placement +
both-safety-copies.
7. CI Postgres pg16 -> pg18 (match docker-compose).
8. jsonb decode seam: shared `parseJsonbValue(value, guard)` in database/utils.ts
holds the legacy double-encoding self-heal in one place; parseToolAllowlist /
parseModelConfig keep only a type-guard.
Verified: server build + 124 unit + 15 integration; mcp 311; prettier clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Admins can now give each EXTERNAL MCP server a free-text instruction ("how/
when to use this server's tools") that the agent receives in its SYSTEM
PROMPT next to the tool descriptions — porting the built-in SERVER_INSTRUCTIONS
idea to admin-configured servers. Trusted, admin-authored text (like a system
prompt); NON-secret, so unlike headersEnc it IS returned in views/forms.
- Migration: nullable `instructions text` on ai_mcp_servers (old rows = null =
no guidance). Table type + repo insert/update (blank/whitespace -> null via
blankToNull). DTO `@MaxLength(4000)`. Service threads it through
McpServerView/toView.
- mcp-clients: `McpServerInstruction { serverName, toolPrefix, instructions }`
threaded through the toolset/cache/lease. Guidance is built ONLY for a server
that actually connected AND contributed >=1 callable tool (the allowlist may
filter all of them out) AND has non-blank text — so a guide never appears for
tools the agent cannot call. Cached with the toolset, so an edit is picked up
next turn via the existing CRUD cache invalidation.
- System prompt: `buildMcpToolingBlock` renders an <mcp_tooling> block INSIDE
the safety sandwich (after context, before the trailing SAFETY_FRAMEWORK) so
it informs tool choice but cannot override the rules; each section is headed
by the server's `prefix_*` namespace. Empty/blank -> block omitted. The
caller (ai-chat.service) now builds the external toolset BEFORE the prompt and
passes external.instructions; client-handle lifecycle (close-once) unchanged.
- Client: instructions field in types + a Textarea (autosize, maxLength 4000)
in the MCP-server form with a namespace-prefix hint; i18n (en/ru).
Tests across every layer (prompt block placement + both SAFETY copies; view
blank->null; buildEntry includes guidance only for connected+>=1-tool+non-blank;
DTO MaxLength; repo + integration round-trip; service wiring). Delegated impl
reviewed (APPROVE); applied the import-type follow-up.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
PR #172 fixed the jsonb double-encoding for `tool_allowlist` but the same
class of bug, and the same re-derived workaround, remained elsewhere.
1. model_config (agent roles): jsonbObject still used the buggy `::jsonb`
bind, so `ai_agent_roles.model_config` round-tripped as a jsonb STRING
SCALAR. The read-path `typeof === 'object'` check then failed and the
model override was SILENTLY dropped (role fell back to the default model).
Fixed to `::text::jsonb` and added `parseModelConfig` + `normalizeRow` so
every read self-heals already-corrupted rows (no migration).
2. Centralized the write workaround as `jsonbBind()` in database/utils.ts —
one implementation with one explanation of the quirk — replacing the
per-repo `jsonbArray` (mcp) and `jsonbObject` (roles).
3. Integration coverage (the fix is a DB round-trip a unit test cannot see;
the read-side parser MASKS a write regression): new
ai-mcp-server-repo.int-spec asserts `jsonb_typeof(tool_allowlist)='array'`
after insert + heals a seeded string-scalar row; ai-agent-roles-repo
int-spec gains the same for `model_config` (`'object'` + heal).
4. Updated the stale `ai-mcp-servers.types.ts` comment (the driver returns a
JSON string for legacy rows; the repo normalizes every read).
5. Fail-open logging: a corrupt tool_allowlist degrades to "no restriction"
(agent gets ALL tools) — normalizeRow now warns (server id only, never
contents) so the silent widening leaves a trace.
6. Simplified parseToolAllowlist (normalize the string once, then a single
array-of-strings check) — identical behaviour, all 12 cases still pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Review caught a real race: onStepFinish fires `updateStreaming()` fire-and-
forget (not awaited), so the FINAL step's streaming UPDATE and the terminal
`finalizeAssistant` UPDATE run as two concurrent statements on different pool
connections — commit order is not guaranteed. If the late streaming update
lands AFTER finalize, the completed row is clobbered back to status='streaming'
with no usage/finishReason, and the next startup sweep then mis-marks the
finished turn 'aborted'. Green unit/integration tests don't reproduce a
cross-connection race.
Fix: scope the per-step update with `onlyIfStreaming` → SQL `WHERE
status='streaming'`. Once finalize has set a terminal status the late update
matches zero rows and no-ops, regardless of commit order; finalize runs
unguarded so it always wins. A cheap `if (finalized) return` short-circuit
avoids most wasted queries, but the SQL guard is the authoritative fix (the
flag can be set after a query is already in flight).
Integration test: finalize to 'completed', then a late onlyIfStreaming update
is a no-op — status/content/usage preserved.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The chat lived in inconsistent paradigms (in-memory stream + client export vs.
DB-as-context), which made export flaky and lost the assistant answer if the
process died mid-turn. Make the DB the single source of truth.
A. STEP-GRANULAR DURABILITY (server)
- ai_chat_messages gains a nullable `status` column (migration; NULL = legacy =
completed). The assistant row is now INSERTED UPFRONT as `status:'streaming'`
and UPDATEd on every onStepFinish with all finished steps (text + tool calls +
tool RESULTS), then finalized once to completed/error/aborted on the terminal
callback. So a process death mid-turn keeps every finished step; a startup
sweep (OnModuleInit → sweepStreaming) flips any dangling 'streaming' row to
'aborted'. The write path no longer depends on a live socket.
- Pure exported `flushAssistant(steps, inProgressText, status, extra?)` builds
the persist payload (metadata.parts byte-identical to the old builder), so a
future background worker can call the same path. AiChatMessageRepo gains
`update`, `sweepStreaming`, and `findAllByChat`.
- consumeStream drain, external-MCP client close-once, SSE heartbeat preserved.
B. SERVER-SIDE EXPORT
- New pure `chat-markdown.util.ts` renders Markdown from DB rows ONLY (server
port of the client builder). Because A persists the in-progress row, the
export now includes an interrupted turn up to its last finished step (flagged
"still generating"). `POST /ai-chat/export` (owner-gated via assertOwnedChat,
workspace-scoped) returns it; `lang` accepts a full client locale tag
('en-US'/'ru-RU') and is normalized server-side (normalizeLang) — a strict
@IsIn(['en','ru']) DTO rejected the real client's i18n.language with a 400,
caught in real-browser testing.
- Client: handleCopy calls the endpoint; `canExport = !!activeChatId`. The whole
liveThreadRef/liveStateRef/onLiveContentChange/hasLiveContent hybrid (and the
client chat-markdown util + test) is removed — the server is now authoritative.
Tests: flushAssistant unit (status shapes + parts parity), chat-markdown.util
unit (incl. legacy NULL-status + interrupted note + ru + normalizeLang locale
tags), controller export wiring + owner-gate, integration update/sweepStreaming.
Verified: server build + 318 ai-chat unit + 3 integration; client tsc + 157
ai-chat unit; and END-TO-END in a real browser — a chat turn persists mid-stream
and the Copy button exports the DB-sourced markdown (showing the in-progress
row), HTTP 200 after the locale fix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Opening the edit form for an MCP server that has a saved tool allowlist crashed
the whole settings page (`TypeError: Ke.map is not a function` in Mantine) — and,
worse, the allowlist was silently NOT enforced. Both stem from one root cause:
the `tool_allowlist` jsonb column round-trips as a JSON STRING, not an array.
Root cause: `jsonbArray` bound `JSON.stringify(value)` (already a JSON string)
straight to a `::jsonb` cast. node-postgres infers the param type as jsonb and
JSON-stringifies it a SECOND time, so the column stored a jsonb STRING SCALAR
(`"[\"a\"]"`, jsonb_typeof = string) instead of an array. On read the driver
hands back the JS string `'["a"]'`. Then:
- the edit form's TagsInput called `.map` on a string -> page crash;
- mcp-clients did `Array.isArray(allow)` -> false for a string -> fell through
to "no restriction" and exposed ALL of the server's tools.
Fix (both verified on the stand):
- Write: `jsonbArray` casts `::text::jsonb` so the param is bound as text (sent
verbatim) and parsed into a real jsonb array. New rows now store
jsonb_typeof=array.
- Read: `normalizeRow` runs every fetched row through `parseToolAllowlist`, which
returns `string[] | null` for both shapes (already-array passes through; a JSON
string is parsed; null/invalid -> null). This REPAIRS existing double-encoded
rows on read, so the UI and the allowlist enforcement work without a data
migration. Applied in findById / listByWorkspace / listEnabled.
- Client: defensive `Array.isArray(...) ? ... : []` guard in the form so a bad
shape can never take the settings page down again.
Tests: ai-mcp-server.repo.spec (8 cases for parseToolAllowlist — array, the
JSON-string read, null, empty, non-array json, unparseable, non-string elements,
non-string primitive). mcp-servers-to-view + mcp-namespacing still green.
Verified live: an old double-encoded row now reads as an array; a newly created
server stores jsonb_typeof=array.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Each chat row in the AI-chat history now shows a dimmed second line with
how long ago the chat was created and the document it was created in
("N ago / <document>", or "No document" when started outside a page).
Server:
- New migration: nullable ai_chats.page_id (FK pages.id, ON DELETE SET NULL).
- Capture the origin page at chat creation from the client-supplied openPage,
but validate it first: it must be a real page in the same workspace that the
user may read (PageAccessService.validateCanView), else null. This keeps the
"openPage.id is attacker-controllable but harmless" invariant - preventing a
cross-workspace/cross-space page-title leak and a post-hijack FK crash.
- findByCreator left-joins pages (scoped by workspace, defense-in-depth) and
returns pageTitle.
Client:
- IAiChat gains pageId/pageTitle; ConversationList renders a ChatMetaLine
(useTimeAgo + origin document) as a dimmed second line.
- Add i18n key "No document" (en-US, ru-RU).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Unit tests for the safety-critical paths: crypto secret-box (round-trip,
tamper detection, wrong key), the SSRF guard (blocked ranges + DNS-rebinding),
the ai-chat tools service, the page-embedding repo, and the
assistant-parts/serialization helpers. Those server helpers (assistantParts,
rowToUiMessage, serializeSteps) are exported ONLY for the tests — no runtime
change.
Also: keyboard a11y on the chat history header and conversation rows
(role/tabIndex/Enter+Space), and DRY refactors that move shared logic into one
place (isToolPart -> tool-parts util; buildInitialValues in the MCP form).
The behaviour-changing edits that previously rode along in this commit are
split out into the following two commits, per review.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Release-cycle test audit found the role feature's security-critical paths
untested. Adds real unit tests (against the actual functions):
- resolveRoleForRequest invariants: role comes from chat.roleId not body.roleId
(no per-turn swap), lookup scoped to workspace.id, disabled/soft-deleted role
-> null, new-chat uses body.roleId, stale chatId falls back.
- CASL admin gate: non-admin create/update/delete -> Forbidden and service not
called; admin delegates with workspace.id; list() is member-reachable.
- roleModelOverride: unknown driver dropped (never reaches getChatModel's
throwing default), valid override passes through, blanks ignored.
- getChatModel override success path (cross-driver fetch + decrypt; chatModel-
only reuse), and service update/remove cross-workspace 'not found' guards +
modelConfig tri-state.
Tiny fix: findByCreator badge left-join now also requires enabled=true, so a
disabled role (downgraded to universal by resolveRoleForRequest) no longer shows
a misleading chat-list badge.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Reusable, workspace-shared agent roles for the built-in AI chat. A role is
a named persona (system-prompt instructions) + optional model override; a
chat is bound to a role at creation and applies it every turn.
Backend:
- migration 20260620T120000: ai_agent_roles table + ai_chats.role_id
(FK ON DELETE SET NULL); hand-merged types into db.d.ts/entity.types.ts
(db.d.ts is hand-curated here, full codegen would clobber it).
- core/ai-chat/roles: CRUD module. list = any workspace member; create/
update/delete = admin (Manage Settings ability, like ai-settings/mcp).
All repo queries scoped by workspace_id; soft-delete (deleted_at).
- buildSystemPrompt gains roleInstructions: role REPLACES the persona base
(admin prompt / DEFAULT_PROMPT) but SAFETY_FRAMEWORK + context are always
still appended.
- stream(): role resolved from ai_chats.role_id for existing chats (never
the request body -> no per-turn role swap); body.roleId only on creation.
Disabled (enabled=false) and soft-deleted roles fall back to universal.
- getChatModel(workspaceId, override): role model_config can swap model id /
driver; a driver without configured creds throws 503 with a clear message
naming the driver+role, resolved BEFORE response hijack.
Client:
- new-chat role picker (enabled roles only, default Universal assistant),
roleId sent only on the first message; role badge (emoji+name) in the chat
header and conversation list; admin Agent-roles management section in
Settings -> AI (add/edit/delete, MCP-form pattern).
Tests: ai-chat.prompt.spec (role layering + safety always present, incl.
jailbreak); ai.service.spec (override on unconfigured driver -> 503).
Implements docs/ai-agent-roles-plan.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add push-to-talk voice dictation that transcribes recorded audio on the
server via the workspace's OpenAI-compatible AI provider (Whisper /
gpt-4o-transcribe / self-hosted whisper), then inserts the text.
Backend:
- New `stt_api_key_enc` column + migration; STT creds parity with chat/
embeddings (sttModel/sttBaseUrl/sttApiKey, write-only key, fallbacks to
chat baseUrl/key). Both provider whitelists updated (service + repo).
- AiService.getTranscriptionModel + AiTranscriptionService.
- Gated POST /ai-chat/transcribe (dictation flag → 403, JWT + workspace
scope + throttle, 25MB cap, MIME whitelist, never logs audio/key).
- New `settings.ai.dictation` workspace flag (DTO + service + audit).
Frontend:
- Wire up the Voice/STT settings card (model/base URL/key) and the
Voice-dictation toggle.
- New `features/dictation`: useDictation (MediaRecorder state machine),
MicButton, transcribe service; integrated into the chat composer and a
new editor-toolbar dictation group, both gated by ai.dictation.
Improve agent RAG quality with three changes, plus a roadmap doc for the rest.
- Indexer: prefix each chunk with its heading path ("Page > H1 > H2"), built by
walking the ProseMirror JSON (heading nodes) so a `#` inside a fenced code block
is never mistaken for a heading. Falls back to plain-text chunking on any error.
buildChunkRows: drop indexOf-against-source offsets (breadcrumb prefixes break
verbatim matching) for a cumulative cursor — offsets are provenance-only.
- Hybrid search: new migration adds a generated `fts` tsvector column + GIN index
to page_embeddings (same english+f_unaccent config as pages.tsv). New
PageEmbeddingRepo.hybridSearch fuses cosine + full-text rankings via Reciprocal
Rank Fusion (k=60, equal weights) in one SQL query at chunk granularity.
- Tools: collapse semanticSearch + searchPages into one hybrid `searchPages` tool
with a query-rewrite-oriented description; gracefully falls back to the REST
full-text path when embeddings are unconfigured. Access control (space scope +
page-permission post-filter) preserved. Add a query-rewrite hint to the default
system prompt.
- docs/rag-improvements-plan.md: record what shipped and the deferred backlog
(reranker, attachment indexing, eval harness, tuning).
Note: requires a corpus reindex to populate breadcrumbs on existing pages.
The WORKSPACE_CREATE_EMBEDDINGS / WORKSPACE_DELETE_EMBEDDINGS jobs were
enqueued (on AI Search enable/disable) but had no AI_QUEUE handler, so
existing pages were never indexed ("Indexed 0 of N pages") and disabling
never purged embeddings.
- EmbeddingProcessor: handle WORKSPACE_CREATE_EMBEDDINGS (bulk reindex all
live pages) and WORKSPACE_DELETE_EMBEDDINGS (purge workspace embeddings)
- EmbeddingIndexerService: add reindexWorkspace() (skips when embeddings
unconfigured; per-page error isolation) and removeWorkspace()
- PageRepo.getIdsByWorkspace(), PageEmbeddingRepo.deleteByWorkspace()
- AiSettingsService.reindex() + admin-only POST /workspace/ai-settings/reindex
- Frontend: "Reindex now" button, service call and mutation
- Stable per-workspace jobId with remove-before-add so a stale job can't
block future reindexes; cancel the delayed purge on enable/reindex so it
can't wipe freshly-built embeddings
Per-workspace AI provider config previously shared a single base URL and
a single API key between the chat model and the embedding model. Add
dedicated, optional embedding endpoint/token that fall back to the chat
values when empty, preserving backward compatibility.
- db: new migration adds nullable `embedding_api_key_enc` to
`ai_provider_credentials`; chat key stays in `api_key_enc`
- repo: add `upsertEmbeddingKey` / `clearEmbeddingKey` (on-conflict
touches only its own column, so chat/embedding keys never overwrite)
- ai-settings.service: store non-secret `embeddingBaseUrl`; resolve()
applies fallback (embeddingBaseUrl || baseUrl; embedding key || chat
key); getMasked() exposes raw `embeddingBaseUrl` + `hasEmbeddingApiKey`,
never the key; update() handles the embedding key write-only
- ai.service: getEmbeddingModel() builds openai/gemini/ollama with the
embedding-specific URL/key; chat path unchanged
- client: new "Embedding base URL" and "Embedding API key" fields with
fallback hints and a clear-key action
Requires running the DB migration on deploy.
Display "Indexed N of M pages" on the AI provider settings page so admins
can see how much of the wiki is covered by vector-RAG semantic search.
- page-embedding.repo: add countIndexedPages() — distinct non-deleted pages
that have stored embeddings in the workspace
- page.repo: add countByWorkspace() — total non-deleted pages
- ai-settings.service: compute both counts in getMasked() (Promise.all) and
return them with the masked settings; inject PageEmbeddingRepo + PageRepo
- MaskedAiSettings / IAiSettings: add indexedPages + totalPages
- ai-provider-settings: render a dimmed coverage line under "Embedding model"
- i18n: add the "Indexed {{indexed}} of {{total}} pages" key (en-US, ru-RU)
so the v6 hook stops re-creating its store every render on a new chat
(which wiped the optimistic user message + streamed deltas, so nothing
showed until the turn finished). Also send X-Accel-Buffering:no + flushHeaders.
- context: client sends the currently-open page {id,title}; the system prompt
tells the agent which page 'this page' refers to (it reads it via its
CASL-scoped getPage tool; id is prompt-context only, no server-side fetch).
- embeddings: make page_embeddings.embedding dimension-agnostic (drop the
HNSW index + ALTER to vector), remove the hard 1536 guard, filter search by
model_dimensions — so 3072-dim (and any) models index instead of being
skipped. Seq-scan <=> search (wiki scale); existing pages reindex on next edit.
- openai provider: use .chat() (Chat Completions) instead of the default callable
(Responses API), which gateways reject on multi-turn -> 400.
- updateAiProviderSettings: assemble settings.ai.provider via jsonb_build_object
with ::text-cast bound params + jsonb_typeof self-heal (postgres.js was
double-encoding it into an array; the ::text cast avoids 'could not determine
data type of parameter').
- chat agent: drop the hard maxOutputTokens cap (truncated complex tool calls);
keep a tiny cap only on the test-connection ping.
- testConnection + chat stream: surface the real provider error (statusCode+message)
to logs and the UI instead of generic masks; never log the API key.
- chat UI: typing indicator, incremental streaming render, tool 'running' status, Stop.
Also bundled (prior uncommitted ai-chat work):
- history 'AI agent' provenance badge; vector RAG (pgvector image + page_embeddings
+ AI_QUEUE indexer + space-scoped semanticSearch); external MCP servers backend
(@ai-sdk/mcp client, SSRF IP-pinning, encrypted headers, admin CRUD/Test);
yjs duplicate-instance fix via pnpm patch (single CJS instance server-side).
WIP checkpoint of the gitmost AI-chat backend (plan stages A + B1 + B3a).
The agent acts under the requesting user's JWT (Docmost CASL enforces page
access); the external service-account /mcp endpoint is untouched.
LLM provider config (A2-A4):
- integrations/crypto: AES-256-GCM SecretBoxService (key derived from APP_SECRET,
per-record salt/iv; clear error on rotation instead of crashing).
- ai_provider_credentials table/repo/types: encrypted API key stored outside
workspace settings/baseFields, write-only (never returned by any endpoint).
- integrations/ai: per-workspace AI SDK v6 provider driver (openai/gemini/ollama),
admin-gated GET(masked)/PATCH(write-only key)/Test endpoints; settings.ai.provider
holds non-secret config incl. systemPrompt. Removed unused AI_* env getters (DB is
the single source of truth).
Chat module (A1, A5-A8):
- ai_chats/ai_chat_messages repos (workspace-scoped, soft-delete, tsv never selected).
- core/ai-chat: CRUD + POST /ai-chat/stream (Fastify hijack + AI SDK v6
pipeUIMessageStreamToResponse, abort on disconnect, persist user/assistant msgs).
- Agent loop: streamText + stepCountIs(8); read tools searchPages/getPage via a
per-request DocmostClient over loopback REST under the user's minted access token.
- Gate settings.ai.chat (+ 503 when provider unconfigured); buildSystemPrompt with a
non-removable safety/anti-prompt-injection framework. Per-user rate limit.
Per-user auth (B1):
- @docmost/mcp DocmostClient gains an additive getToken variant (carry a user JWT,
re-fetch on 401) and exports DocmostClient; the email/password service-account path
(external /mcp, stdio) is unchanged.
Agent-edit provenance backbone (B3a):
- Migration: pages/page_history (last_updated_source, last_updated_ai_chat_id) and
comments (created_source, ai_chat_id, resolved_source).
- Signed actor/aiChatId claim in the collab token; onAuthenticate propagates it,
onStoreDocument writes it with a sticky agent marker, saveHistory copies it.
Migrations auto-run on boot (additive). Write tools, frontend, RAG and external MCP
servers are not in this checkpoint.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
